摘要
金属伪影校正对提高CT图像清晰度具有重要意义。针对金属伪影校正研究中伪影消除不彻底、组织结构缺失等问题,提出一种基于残差编解码网络的金属伪影去除(RED-CNN-MAR)方法。首先,使用RED-CNN网络实现由金属伪影图像到无金属伪影图像的端到端映射,在卷积层之后引入BN层提高网络训练精度和加快收敛速度;并且将原始图像、线性插值(LI)图像、射束硬化校正(BHC)图像作为RED-CNN网络的三通道输入,以融合不同校正方法的优势。接着,对该网络的输出图像在投影域进一步做组织处理;最后利用滤波反投影重建得到校正后的无金属伪影图像。经实验分析,经过RED-CNN-MAR方法校正后的图像RMSE减小了0.000 7,PSNR和SSIM分别提高了0.59 dB、0.002 8。实验结果表明,该方法可以有效地抑制金属伪影,重建出清晰的结构细节。
Metal artifact reduction is of great significance to improve the clarity of CT images. In this domain, some problems include incomplete artifact removal and miss of organizational structure. To address these issues, proposes a method of metal artifact reduction based on the residual encoder-decoder network(RED-CNN-MAR). Firstly, the RED-CNN network is used to realize the end-to-end mapping from the metal artifact image to the metal artifact-free image. The BN layer is utilized after the convolutional layer to improve the training accuracy of the network. Meanwhile, the speed of convergence is enhanced. To integrate the advantages of different correction methods, the original image, linear-interpolation images and beam-hardingcorrection images are used as the three-channel input of the RED-CNN network. Secondly, the output image of the network is further processed in the projection domain. Finally, the corrected image without metal artifact is reconstructed by the filtering back projection algorithm. The RMSE of the image corrected by the RED-CNN-MAR method is reduced by 0.000 7. PSNR and SSIM are improved by 0.59 dB and 0.002 8, respectively. Experimental results show that the proposed method can effectively suppress metal artifactand reconstruct clear structural details.
作者
马燕
余海军
钟发生
刘丰林
Ma Yan;Yu Haijun;Zhong Fasheng;Liu Fenglin(Key Lab of Optoelectronic Technology and Systems,Ministry of Education,Chongqing University,Chongqing 400044,China;Engineering Research Center of Industrial Computer Tomography Nondestructive Testing,Ministry of Education,Chongqing University,Chongqing 400044,China;State Key Lab of Mechanical Transmission,Chongqing University,Chongqing 400044,China)
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2020年第8期160-169,共10页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(61471070)
国家重大仪器开发专项(2013YQ030629)资助
关键词
金属伪影校正
深度学习
残差网络
编解码器
metal artifact reduction
deep learning
residual network
encoder-decoder
作者简介
马燕,2018年毕业于曲阜师范大学获得学士学位,现为重庆大学硕士研究生,主要研究方向为仪器科学与技术、金属伪影校正。E-mail:mayan@cqu.edu.cn;通信作者:刘丰林,分别在1990年、1993年和2009年于重庆大学获得学士学位、硕士学位、工学博士学位。现为重庆大学研究员、博士生导师,主要研究方向为工业CT技术与系统、机械电子技术。E-mail:liufl@cqu.edu.cn
